VOLSHi / ERIE

Retrieval without hallucination.

ERIE — the Epistemic Retrieval & Investigation Engine. When a bot asks about your user, everything that comes back carries its provenance. No invented answers. No confident fiction. And when the evidence isn't there, she says so — explicitly.

5,201 tests passing · Python · stdlib-only core
How she works

A question goes in. An epistemic result comes out.

ERIE sits between the question and the answer. She derives what would count as an answer, plans the evidence she'd need, retrieves it, and judges whether it's sufficient — sufficient, insufficient, contradictory, stale, or ambiguous. Then she reports.

01

Target

The question becomes a target with satisfaction criteria — what an answer would have to establish before anyone gets to speak.

02

Evidence

Retrieval runs through an adapter over your data pool. Every candidate is bound back to the criteria it supports, with its provenance attached.

03

Verdict

Sufficiency is judged, not assumed. Answered — with the full chain of custody. Or cannot-answer — with the gaps preserved, not paved over.

The shape of every answer

Provenance isn't a feature. It's the output.

Example — the structure ERIE returns for every question. Real engine output shape; the data below is illustrative.

erie.ask — example
Q: "Why was acct-002's name changed?"
→ answered
evidence: k-record interval [k2..live] · cust_name = "Sofia Bennett"
provenance: katie.entries row 47 · reason "married" · authority 89%
gaps: none — criteria satisfied

Q: "What is the customer's mother's maiden name?"
→ cannot-answer
gaps: no evidence in the pool addresses this · refused to invent
The honesty boundary

She won't invent. She won't vouch for the pool, either.

ERIE guarantees the chain of custody — source to evidence to answer — and she would rather say cannot-answer than fabricate. What she does not do is independently verify that your data pool tells the truth; that's the pool's job. In our stack the pool is KATIE's K-record — append-only, every entry bound to its moment — so the ground she stands on is solid. We say exactly this much, and no more.

The trifecta

One engine of three.

01

KATIE

Watches. Every belief, every decision, every handoff — issued a receipt and written to the K-record.
03

ELLE

Persists. The external learning loop — experience turned into durable lessons that survive model replacement.
⊕

ARCHE

The architecture they run on — the emulator of how humans function, with memory as part of the design.

Ask. Get the truth or get told.

ERIE is the retrieval layer for systems that can't afford confident fiction. Grounded answers, full provenance, honest gaps.

Talk to us